From Design and Construction to Discovery: Machine Learning Algorithms for Particle Physics Triggering and Tracking with GPUs and FPGAs in Malaysia

Malaysia is currently in the midst of its own Data Analytics (DA) revolution. Through governmental policy, it supports both academic research and industrial avenues to increase Malaysia's capacity for DA and High Performance Computing (HPC). In the proposed project, which will run for three years from April 2018, physicists at Imperial College London seek to participate in this effort by working together with leading researchers from Malaysia on the Coherent Muon to Electron Transition (COMET) Experiment.

COMET is designed to investigate charged lepton flavour violation (CLFV) by searching for the as-yet-unseen muon to electron conversion on an aluminium nucleus. This process is not allowed in the Standard Model of particle physics, but has very good sensitivity to Beyond the Standard Model physics. A positive discovery of CLFV would be a Nobel Prize-class result and revolutionise our understanding of the building blocks of the universe. The experiment is currently under construction at the internationally-renowned Japan Proton Accelerator Research Complex (J-PARC) laboratory.

The first phase of the COMET experiment will begin taking data in 2019. This phase is designed to probe this process 100 times better than the current limit, with a second phase reaching a sensitivity that is 10,000 times better. To achieve this sensitivity, we are building the world's most intense muon beam which will deliver over a billion muons per second. Very few of these muons will result in anything that is relevant to muon to electron conversion, and instead will form the unwanted background to our sought-after CLFV signal. Therefore COMET (a search for a miniscule "signal" in the midst of a billion billion mundane "background" occurrences) presents a huge technical challenge which will be met by the use of a combination of modern technologies: these include cutting-edge applications of Field-Programmable Gate Arrays (FPGAs; programmable integrated circuits which allow real-time data algorithms to be implemented) and Graphics Processing Units (GPUs) as accelerative computing components, supercomputing via a cloud-based computing solutions, and Machine Learning (ML) techniques. Whilst these will be employed to allow COMET to pursue its particle-physics goals, these are technologies that are known to offer novel solutions to common problems in a growing number of global DA industries, from communications and networking to financial applications and blockchain research.

COMET's first data-taking runs are scheduled for 2019, in the second year of the project. Over a hundred physicists from 35 institutions in 14 countries currently participate, and top scientists from Malaysia will work with researchers from Imperial on the preparations for the experiment, real data-taking and the physics analysis of the data. The Imperial group designed the software that is being used by COMET and participants will take part in the global collaborative coding effort, and the production of massive amounts of computer simulations that are needed to allow us to study the data, and work with experts in FPGA and GPU programming to produce and test the hardware and software that is needed by the experiment. The project includes attachments in the UK and Malaysia to allow the exchange of skills, as well as extended time at the J-PARC laboratory to build and run the experiment together with our international colleagues.

The project will result in participants taking on valuable roles within the experiment, being trained in modern DA techniques and applying them to this world-leading experiment. In addition to a particle physics measurement that has the potential to be a paradigm-changing discovery, it will provide for a solid basis for future leadership in a broad range of pure and applied research.

The primary physics goal of the Coherent Muon to Electron Transition (COMET) experiment and the technical research goals of this project both represent high-impact research. COMET will make the world's most sensitive measurement of charged lepton flavour violation (CLFV). This search parallels the successful discovery of neutral lepton flavour violation in the neutrino sector, which has opened up a wealth of high-impact research avenues across the globe. To support COMET's search for CLFV, this project will utilise cutting-edge data processing techniques commonly used in industry while pushing the boundaries of these techniques by implementing them on novel computing architectures. This knowledge exchange between industry and academia is designed to fuel the rapid economic and academic growth of Malaysia's data analytics and high performance computing sectors, with the results will be published in academic journals so that the wider industrial and academic communities will benefit.
COMET will search for CLFV, where a positive result for this measurement would be a Nobel-prize level discovery, whilst the absence of the CLFV signal is also a high-impact result, with COMET improving on the current sensitivity by a factor of 10,000. All theoretical extensions to the Standard Model of particle physics would have to accommodate this measurement, while all future CLFV experimental searches would need to surpass this sensitivity. No matter the outcome, COMET's final measurement will be widely published and welcomed.
The first goal of this project will push the boundaries of data analytics and high performance computing in Malaysia and in the larger academic context. In the past decade, machine learning has transformed many data analytic industries. This is largely due to the adoption of Graphics Processing Units (GPUs) as more efficient and more powerful computing elements than Central Processing Units (CPUs). Through a partnership between UK and Malaysian scientists, we will integrate GPUs into COMET's particle trajectory reconstruction algorithms. This will equip Malaysian researchers with the GPU expertise that can grow Malaysia's data analytics capabilities. Publications in physics and technology journals will demonstrate to the global research community the power of adopting GPUs into uniquely-demanding data processing environments.
The second goal will be to explore the potential of Field Programmable Gate Arrays (FPGAs) as a next-generation computing architecture for machine learning. Similarly to the reasons why GPUs are overtaking CPUs as an industry standard in high performance computing, FPGAs use even less power and, unlike GPUs, can offer faster computation times with fixed latencies. We will implement machine learning on FPGAs as key decision-making parts of COMET's high-bandwidth data acquisition system. This will demonstrate to both academia and industry the untapped potential of implementing machine learning on FPGAs.
This project will encourage knowledge exchange between Malaysia and the UK through collaboration on high-impact research. This affords the UK with the opportunity to take part in Malaysia's rapidly expanding data analytics sector while providing Malaysia with necessary expertise to fuel this expansion. The research goals directly support COMET's larger physics goals, which will have a lasting impact on the global physics community and the world's understanding of fundamental physics.
The Imperial College group has been involved in the COMET collaboration since its inception, and members have been central to its management, having being represented on the Executive Board since its establishment. Participants in this project will be fully included in the collaboration and their contributions and outputs will be managed under the same principles of open access, and both academic and technical outputs will be valued highly.